[nightly-eval] Nightly regression: contract_roman_numeral (2026-07-19)

[sunholo-voight-kampff]

Nightly eval REGRESSION: benchmark 'contract_roman_numeral' passed ALL trials in the previous run but failed BOTH trials on 2026-07-19.
Error category: [compile_error]
Model: opencode-qwen3-5-35b-a3b-mxfp8 (local, tiers:smoke,core)
Results: /tmp/nightly_eval_20260719 (prev: /tmp/nightly_eval_20260718_rag_on/agent)
It was solid last night, so this is a fresh solid->broken regression — investigate.

---

Binary info (auto-attached):
ailang version: v0.29.2-421-g81a45f2d8
git commit: 81a45f2

---

Reported by: nightly-eval via ailang messages

[sunholo-voight-kampff]

Closing after a stale-regression audit (2026-07-20, prompted by #417 — thanks @yashpatel1230): this regression did not persist. The benchmark passes in subsequent runs (v0.30.0 baseline, 2026-07-19) and the nightly bot — which re-flags any regression that persists — did not re-flag it in the latest run (2026-07-20). Verdict: transient/model-variance, recovered. Process fix landing today: the mission loop's triage now closes the corresponding issue whenever it refutes or resolves a flagged regression, so stale alarms no longer linger open.